DeepSeek V4: Does China Just Matched GPT-5?

2026-05-22 · Source: Deep Learning on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Advanced, short

Summary

DeepSeek has released V4 Preview (Pro + Flash) under an MIT license, claiming near-frontier reasoning capabilities at approximately 1/6th the cost of leading Western models. The release features a 1M-token context window, enhanced agentic coding, and training on Huawei Ascend 910C chips. V4-Pro, with 1.6T total / 49B active parameters, and V4-Flash, with 284B total / 13B active parameters, both support three reasoning effort modes. Architecturally, V4 introduces Hybrid Attention, reducing per-token FLOPs by 73% and KV cache by 90% compared to V3.2 at 1M context, alongside FP4 precision on MoE weights. Benchmarks show V4-Pro-Max achieving open-source SOTA on SWE-bench Verified and LiveCodeBench/Codeforces, and competing closely with GPT-5.5 and Opus 4.7 on BrowseComp, while trailing on SWE-bench Pro and Terminal-Bench 2.0. Notably, factual recall improved significantly, with SimpleQA Verified leaping from 28 to 55 and FACTS Parametric from 27 to 63.

Key takeaway

For AI Engineers evaluating new LLMs, DeepSeek V4 presents a compelling open-weight option that rivals top proprietary models in specific areas at a fraction of the cost. You should test V4-Flash for most applications, leveraging its API compatibility for tasks like code generation or web agents. Be aware of opaque training data and potential quality differences in creative writing or literary translation, where Anthropic models may still excel.

Key insights

DeepSeek V4 offers near-frontier performance at significantly lower cost and open weights, challenging established proprietary models.

Principles

• Inference economics drive frontier model accessibility.
• Hybrid Attention reduces FLOPs and KV cache usage.
• Open-weight models can achieve SOTA in specific domains.

Method

DeepSeek V4 employs a two-stage post-training pipeline: domain-expert SFT + GRPO-based RL, followed by unified on-policy distillation, after 32T+ token training.

In practice

• Use `deepseek-v4-flash` for most tasks due to speed.
• Integrate with OpenAI ChatCompletions or Anthropic APIs.
• Self-host frontier-adjacent models for on-prem needs.

Topics

• DeepSeek V4
• Large Language Models
• Mixture-of-Experts
• Inference Optimization
• Open-weight Models
• Benchmarking
• Huawei Ascend

Best for: CTO, VP of Engineering/Data, AI Architect, AI Engineer, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• DeepSeek v4 — DeepSeek V4 advances open-weight long-context and agentic coding with a novel architecture and competitive performance.
• DeepSeek v4: the most expected open-source model ever released, and the quietest landing — DeepSeek v4 prioritizes serving cost reduction through architectural innovation over raw benchmark performance.
• Z.ai GLM-5: New SOTA Open Weights LLM — Chinese AI labs are rapidly advancing open-weight models, emphasizing cost-efficiency, long-context, and agentic capabilities.
• DeepSeek STRIKES BACK - Open-Source Model Challenges OpenAI And Anthropic — DeepSeek V4 sets a 1 million token context as standard, democratizing advanced AI capabilities through open-weight MoE models.
• DeepSeek previews new AI model that ‘closes the gap’ with frontier models — DeepSeek V4 models offer competitive performance and unprecedented affordability for large context window MoE architectures.
• As agentic AI pushes rivals to raise prices and cap usage, Deepseek ships a good-enough model for almost nothing — Deepseek's V4 models offer massive context windows and competitive performance at significantly reduced costs via architectural innovation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Deep Learning on Medium.